Customer Support Agent

A vertical-agnostic inbound support pattern: customer dials your support number, the agent identifies them via inbound webhook, looks up their account context, answers tier-1 questions from a knowledge base, opens a ticket if needed, and warm-transfers to a human for anything outside its scope.

What it does

1. Greets the caller by name (via the inbound webhook)
2. Asks "How can I help today?"
3. For known FAQs: answers from the knowledge base (RAG tool)
4. For account-specific issues: opens a ticket via tool and confirms the ticket ID
5. For complex / emotional / out-of-scope issues: warm-transfers to a human
6. Captures the topic, resolution, and CSAT signal

Suggested configuration

Code
# Knowledge base must exist first
kb = client.knowledge_base.from_url(
    url="https://your-help-center.example.com",
    name="Help Center",
)

agent = client.agents.create(
    name="Tier-1 Support",
    brain={
        "provider": "phantom",
        "model": "phantom",
        "instructions": SUPPORT_PROMPT,
        "temperature": 0.4,
    },
    voice={
        "provider": "cartesia",
        "model": "sonic-3",
        "voice_id": "8d8ce8c9-44a4-46c4-b10f-9a927b99a853",
    },
    transcriber={"provider": "deepgram", "model": "nova-3", "language": "en-US"},
    interruption={"enabled": True, "min_duration": 0.4},
    timeout={"enabled": True, "trigger_seconds": 12},
    call_duration={"enabled": True, "max_seconds": 900},  # 15-min cap
    inbound_webhook_url="https://your-app.example.com/neuratel/inbound-context",
    analytics={
        "recording": {"enabled": True},
        "summary": {"enabled": True},
        "success_evaluation": {
            "enabled": True,
            "rubric_type": "pass_fail",
            "criteria": "Did the agent either resolve the issue from KB, open a ticket with confirmation, or warm-transfer to a human?",
        },
        "data_extraction": {
            "enabled": True,
            "fields": [
                {"name": "topic", "type": "string"},
                {"name": "resolution_path", "type": "enum",
                 "values": ["resolved_kb", "ticket_opened", "transferred", "abandoned"]},
                {"name": "ticket_id", "type": "string"},
                {"name": "csat_score", "type": "number"},
            ],
        },
    },
    tools={
        "rag": {
            "enabled": True,
            "knowledge_base_ids": [kb["id"]],
            "top_k": 5,
            "score_threshold": 0.7,
        },
        "hangup": {"enabled": True, "keywords": ["bye", "goodbye", "thanks bye"]},
    },
    transfer={
        "enabled": True,
        "mode": "warm",
        "destinations": [{
            "name": "Tier-2 Support",
            "number": "+1...",
            "description": "Escalate billing disputes, account changes, technical issues beyond FAQ",
            "keywords": ["speak to human", "supervisor", "billing dispute", "account change"],
        }],
        "warm_intro_template": "I have {{first_name}} on the line — they're asking about {{topic}}. Account tier: {{tier}}. They've been on with me for {{call_duration_secs}} seconds.",
    },
)

# Wire to inbound number
client.phone_numbers.assign(phone_number_id="num_...", agent_id=agent["id"])

Inbound webhook

Configure your endpoint to look up the caller and return context — see Inbound variable injection for the full webhook contract. Minimum payload to return:

Code
{
  "dynamic_variables": {
    "first_name": "Alice",
    "tier": "scale",
    "open_tickets_count": "1",
    "is_repeat_caller": "true"
  }
}

Prompt template

Code
You are a friendly support representative for {{company_name}}.

Caller context (from your CRM):
- Name: {{first_name}}
- Tier: {{tier}}
- Open tickets: {{open_tickets_count}}
- Repeat caller: {{is_repeat_caller}}

GREETING:
- If repeat caller with open tickets: "Hi {{first_name}}, thanks for calling
  {{company_name}}. I see you have {{open_tickets_count}} open ticket — is
  this related, or something new?"
- Otherwise: "Hi {{first_name}}, thanks for calling {{company_name}}. How can
  I help you today?"

DECISION TREE:
1. Listen to their question. Use the RAG tool to look up the topic.
2. If RAG returns a confident answer (relevant, on-topic):
   - Summarize in your own words. Don't read the help center verbatim.
   - Confirm: "Did that answer your question?"
3. If they need an account change or backend action (refund, plan change,
   data export request):
   - Don't promise the action yourself. Open a ticket with the open_ticket tool.
   - Confirm the ticket ID with them. Tell them when to expect a response.
4. If they're frustrated, the issue is complex, or they ask for a human:
   - Warm-transfer to Tier-2: "Let me get a specialist on for you — I'll fill
     them in before connecting you."
5. Before ending: "On a scale of 1 to 5, how would you rate this call?"

GUIDELINES:
- Don't make up facts. If RAG doesn't return an answer, say so:
  "I don't have a confident answer on that — let me get a specialist."
- Never give pricing without RAG confirmation.
- Acknowledge frustration before redirecting: "I'm sorry that happened — let
  me see what I can do."
- Be brief. One short sentence at a time.
- If the caller is in distress (medical emergency, threat), say "Please hang up
  and call your local emergency services" and end the call.

Integration tips

• Knowledge base freshness: re-sync your help-center KB on a schedule (from_url is incremental). Outdated KB content is the #1 cause of "hallucinated" answers.
• Ticketing integration: build a custom MCP tool wrapping your ticketing system. The agent calls open_ticket(category, summary, severity) and returns the ticket ID via variable_updates.
• Sentiment routing: the analytics rubric captures CSAT, but for real-time escalation tie a low-sentiment threshold to the transfer tool's keywords.
• Out-of-hours behavior: configure a separate agent for nights/weekends that skips the transfer step and offers callback scheduling instead.

Related

• Inbound variable injection — caller context lookup
• Knowledge base — RAG setup
• Call transfer — warm transfer mechanics
• Tool variable writes — capturing ticket IDs from tools